Analytics Engineering & NLP: Churn Diagnostics and Financial Risk in SaaS

1. The Executive Challenge: The Churn "Black Box"

GenSoft, a tech company operating under the SaaS (Software as a Service) model, reached the 20,000 active customers milestone. However, rapid growth brought a severe side effect: a critical spike in churn rates and a flood of negative feedback across public channels and support tickets.

The board needed quick answers to three fundamental questions that defined the business's future:

The core issue wasn't a lack of data, but the inability to process it. With a historical database of over 45,000 free-text reviews (natural language), the board was flying blind. Reading and categorizing this volume manually was humanly impossible. My mission as an Analytics Solutions Architect was to decode this mass of unstructured data and build a prescriptive Data Product.

2. The Architectural Hurdle: API and Network Limitations

The obvious solution would be connecting the database to an AI API (like OpenAI) to perform Sentiment Analysis and Topic Modeling. However, in the real world of Data Engineering, we hit critical roadblocks:

• Cost and Rate Limits: Free APIs (like Hugging Face's Inference API) have strict limits. Sending 45,000 simultaneous requests would result in an immediate IP block (Error 429 - Too Many Requests).
• Power BI Network Isolation: The Python script execution environment inside Power Query has severe DNS resolution restrictions, preventing direct HTTP calls via the requests library.

To bypass these limitations, I designed a Hybrid Architecture. Python was used strictly for local in-memory processing (Statistical Sampling), while the M language (Power Query) took over the network request responsibility via Web.Contents, bypassing the internal firewall.

3. The Statistical "Hack": Cochran's Formula

To avoid exhausting API requests and optimize model refresh time from hours to minutes, I applied a fundamental Data Science concept: you don't need to drink the ocean to know the water is salty.

I utilized Cochran's Formula with Finite Population Correction. Assuming a 95% Confidence Level and a 5% Margin of Error, the Python script reduced the 45,218-row database to a stratified sample of just 380 records. This slashed cloud processing volume by over 99% while keeping mathematical accuracy intact.

import pandas as pd
import numpy as np

df = dataset.copy()

# 1. COCHRAN'S FORMULA PARAMETERS
Z = 1.96          # 95% Confidence Level
p = 0.5           # Expected proportion (maximum variance)
e = 0.05          # 5% Margin of Error
N = len(df)       # Total population (45,218)

# 2. SAMPLE CALCULATION WITH FINITE CORRECTION
n0 = (Z**2 * p * (1-p)) / (e**2)
n = int(np.ceil(n0 / (1 + (n0 - 1) / N)))

# 3. STRATIFIED SAMPLING (Maintaining original score proportions)
sample = df.groupby('Original Score', group_keys=False).apply(
    lambda x: x.sample(n=max(1, int(np.round(len(x) / len(df) * n))), random_state=42)
).reset_index(drop=True)

4. Generative AI Integration (Hugging Face)

With the sample perfectly isolated, I built a custom M Language function to query Hugging Face's modern router (router.huggingface.co).

I used the nlptown/bert-base-multilingual-uncased-sentiment model, optimized for multiple languages. The M function sends the text, receives the classification (1 to 5 stars), and applies GenSoft's business rules to translate the output into "Positive", "Neutral", or "Negative".

6. Data Modeling and Decision Intelligence (DAX)

The model was structured in a rigorous Star Schema. To ensure dashboard internationalization (i18n) without duplicating pages, I implemented a 100% bilingual architecture using Disconnected Tables and Field Parameters.

The major differentiator of this project for the board was the creation of a Financial Simulator (What-If Parameters). Instead of just showing past Churn, the dashboard allows the C-Suite to adjust plan prices (Basic, Premier, Deluxe) in real-time to forecast the financial impact of dissatisfaction.

For the calculation to be mathematically sound, the DAX measure had to perform Upscaling (Statistical Extrapolation), multiplying the sample result by the population factor.

Simulated_Revenue_at_Risk = 
// 1. Calculate Extrapolation Factor (Population / Sample)
VAR Extrapolation_Factor = DIVIDE(MAX(fReviews[_Total_Population]), MAX(fReviews[_Sample_Size]), 1)

// 2. Calculate simulated revenue iterating over the plans table
VAR Simulated_Sample_Revenue = 
    SUMX(
        dPlans,
        VAR Qty_Negative_Reviews = 
            CALCULATE(COUNTROWS(fReviews), fReviews[fnGetSentiment] = "negative")
        VAR Simulator_Value = 
            SWITCH(
                dPlans[Plan Type],
                "Basic", [Basic Plan Simulator Value],
                "Premier", [Premier Plan Simulator Value],
                "Deluxe", [Deluxe Plan Simulator Value], 0
            )
        RETURN Qty_Negative_Reviews * Simulator_Value
    )

// 3. Project the real value for the entire base
RETURN Simulated_Sample_Revenue * Extrapolation_Factor

7. Executive Answers and Action Plan

Semantic analysis, the interactive Treemap, and data modeling accurately answered the board's questions, shifting GenSoft's strategic direction:

This case demonstrates that the true value of Analytics doesn't lie in plotting charts, but in orchestrating Data Engineering, Statistics, and Artificial Intelligence to protect company revenue and guide executive decision-making.